Multiple-model tracking with fixed-lag smoothing using imprecise information
A new multiple-model filter for target tracking has been developed and performed well in this thesis. The procedure of the new multiple model (MM) filter has no compromises between non-manoeuvre and manoeuvre, and between manoeuvres except in the ambiguous cases. The operation of the new MM filter is simple like VD filter but with no need for reconstruction of manoeuvre, and the new MM filter also considers all kinds of motions like interacting multiple model (IMM) filter but with a small number of models and significantly reduced computational load. The scheme of the new multiple-model tracking filter consists of manoeuvre detection, construction of manoeuvre filters, construction of safeguard filter, and filter selection. The performance of the proposed tracking filter mainly relies on manoeuvre detection, construction of manoeuvre model filter, and construction of safeguard model filter. In order to improve the tracking, several manoeuvre detection methods have been developed. One of the manoeuvre detection methods is to test the statistic of normalised optimally squared smoothed accelerations, and gives quicker detection than classical manoeuvre detection by testing normalised squared-innovation statistic, in spite of the smoothing lag. This thesis suggests that the manoeuvre be detected by testing the changes of the statistic of normalised squared innovations to give effective manoeuvre detection, based on Chen and Norton's (1986) detection by testing rapid parameter changes. The thesis also modifies Weston and Norton's (1997) change detection with the fixed-lag smoothing instead of the fixed-interval smoothing used by Weston and Norton's method, and obtains more accurate and quicker manoeuvre detection. According to the features of target motion, the target manoeuvres are modelled as straight-line acceleration motion, cross-track acceleration motion, and curvilinear acceleration motion. Thus the manoeuvre model filters can be constructed by these three kinds of motions with a limited number of manoeuvre model filters and reduced computational load. To avoid the risk of the loss of track, a safeguard filter is used in the case of uncertain manoeuvre. The safeguard filter is constructed by combining Singer's (1970) filter and input estimation, to provide at least comparable performance to IMM filter. Further improvement for multiple-model tracking is provided by using the fixed-lag smoothing technique. In comparison with the multiple-model filter alone, the fixed-lag smoothing multiple-model filter provides much better performance (even with fixed lag d=1), and can be implemented in a real time at the costs of a small delay and slight increase in computational load.